This project explores the classical computer vision technique (non-deep learning) of converting 2D images into 3D Reconstruction. By harnessing the power of computer vision methodologies and utilities, I converted two images depicting a castle into a three-dimensional representation. A pivotal aspect of this effort was ascertaining the camera's positions when the photographs were captured, followed by the creation of the 3D visual.
> Highlighted milestones of this endeavor are:-
Retrieving 3D Transformations:
- The objective here was to uncover the 3D transformation (R, T) between dual viewpoints. For two points (P1, P2 in R^3) that represent the same scene in the first and second frames, the relationship is given by (P2 = RP1 + T).
-
Deriving the Essential Matrix:
- Method of Least Squares:
- I utilized the 8-point technique to deduce the essential matrix (E) by way of the SVD decomposition approach.
- RANSAC Technique:
- Depicting Epipolar Lines:
- With the help of the essential matrix, I illustrated the epipolar lines on both snapshots, providing insight into the correlation between corresponding points.
- Method of Least Squares:
-
Deciphering Pose & Constructing 3D:
- Resolving Twisted Pair Ambiguity:
- Investigated the solutions concerning twisted pair ambiguity associated with (E), which gave rise to four potential solutions for the transformation (R, T).
- Point Triangulation:
- Points were triangulated to pinpoint the confluence of rays from the two cameras. From the four potential transformations, I refined the 3D construction by opting for the transformation where the majority of the 3D points appeared in the foreground for both cameras.
- Resolving Twisted Pair Ambiguity:
-
Reprojection Verification:
- As an essential verification step, I juxtaposed points from one camera's original snapshot against the reprojected points from the alternate camera. This process called for the application of the camera's projection blueprint and a deep dive into the transformation's articulation across the two viewpoints.